Recent updates for ATLAS+CMS combination

• Improve discrete minimization to skip the final fit
• Support CMS workspace in xml_ws_combiner.py by both RooSimultaneousOpt and RooSimultaneous
• Speed up log
• Implement discrete minimization for RooSimultaneous
• Bugfix for bbtautau and bbbb

marked the checklist item Improve discrete minimization to skip the final fit as completed

marked the checklist item Support CMS workspace in xml_ws_combiner.py by both RooSimultaneousOpt and RooSimultaneous as completed

marked the checklist item Speed up log as completed

marked the checklist item Implement discrete minimization for RooSimultaneous as completed

added 1 commit

• 329934a4 - Bugfix in unfold_constraints

Compare with previous version

marked this merge request as ready

Hi @clcheng, the CMS bbbb and bbtt workspaces are not supported yet with useCMSOptPDF=False, I'll work on this issue. But for now I think current developments could be merged first.

added 2 commits

• f5240cb1 - signal injection for cls limit
• 0f7923d6 - fix class judgement

Compare with previous version

in order not to change the key order in the limits dictionary, I suggest doing:

if ...
    inj_limit = self.get_limit(asimov_inj_nll, med_limit, 'inj')
else
    inj_limit = 0.

then do

limits['inj'] = inj_limit

Thank you!

I suggest setting mu_inj:Optional[float]=None

then do

if self.mu_inj is not None:

what is the reason for not modifying the global observables here?

Should we? it seems that we didn't do any fit here

And why are you not doing any fit here? I think it depends on whether you are blinded or unblinded.

This option is just a signal injection test, which is different from SM-hypothesis limit setting.

Why in a signal injection test should the nominal NP values always be used?

from my understanding, this Asimov production will use the NPs which is used to construct asimove_data_0, but set POI=1, so it is not always nominal NP, but always asimove_data_0 + signal

ok just realized that you have loaded the global observables and NPs

self.load_conditional_snapshot(self.mu_exp, target=WSArgument.GLOBAL_OBSERVABLE)
self.load_conditional_snapshot(self.mu_exp, target=WSArgument.NUISANCE_PARAMETER)

That is fine then

Did you try to load back the original NPs and global observables once your signal injection is finished? Just to be safe.

I check the case with or without signal injection, and the result is the same. But I can do that

Thanks

added 3 commits

• 5a5e4612 - Improve injection
• deb2315b - set useCMSOptPDF by default
• 48272c5e - Revert "Print log as soon as possible"

Compare with previous version

resolved all threads

added 1 commit

• 34ca2a90 - improvement

Compare with previous version

mentioned in commit ebb6656f

merged